Opus 5.5 fixes what Opus 5 broke
- YT :: https://www.youtube.com/watch?v=jgGyX7MPPVg
- Original title :: Anthropic Actually Fixed Opus
Theo's day-one review of Anthropic's Opus 5.5. After calling Opus 5 a liability, he finds 5.5 to be a genuine recovery: cheaper, faster, far better at following instructions, and dramatically more readable in its prose. He spent a full day pushing it hard and burned roughly 20% of his weekly Claude limit, which he frames as the best value currently available in a $200 coding subscription.
What Anthropic shipped
Anthropic positions Opus 5.5 as performing at Fable 5.1 level on most work at 40% less cost to run than Opus 5. Token prices drop 20% to $4/M in and $20/M out, cache reads get 60% cheaper at $0.20/M, and serving is about 30% faster. Sonnet 5.5 and Haiku 5.5 are teased for the coming weeks — the first Haiku release in a year.
A small but real fix: the reasoning effort level no longer lives at the top of the system prompt, so switching effort mid-session no longer busts the cache.
Anthropic also highlights communication quality as the headline improvement, and Theo agrees. He deleted his anti-slop instructions from CLAUDE.md to see how the model writes unguided and does not plan to put them back. He also notes Opus retains zero-data-retention eligibility, which Fable lacks — enough on its own to make Opus the default coding model inside large enterprises.
Cost and token efficiency
The price cut is not an efficiency win. Opus 5.5 on max burns almost 120k tokens per task versus 73k for Opus 5, 78k for Fable, and 27k for GPT-6 Astra — a roughly 4x efficiency gap against OpenAI. Total per-task cost lands near $6, close to the most expensive model artificial analysis has measured, saved only by the price and cache cuts.
The picture changes sharply below max. On medium it is about $1.34 per task against $3.26 for Astra, and Opus 5.5 sweeps the Pareto frontier for every score above 50. Theo's repeated advice: avoid max entirely. On SkateBench, moving from xhigh to max took reasoning from ~330 tokens per question to almost 5,000 — over 10x the cost for a 1% score change — and he personally lost a six-and-a-half-hour session to a max-effort run looping on a markdown plan. Low is likewise not worth its price; medium through xhigh is the useful band.
Where it is strong and where it is not
Benchmarks are broadly excellent: a five-point jump over Fable 5.1 on the artificial analysis intelligence index, new highs on Terminal Bench 4 and Cursor Bench, often scoring higher on medium than Fable did on max.
Front-end design is the clear weak spot. Theo finds its generated pages competent but sloppy compared to Fable 5.1, which still has better instincts for layout, animation, and taste — worth reaching for on marketing pages and design overhauls. Others disagree: Mia ran a hundred HTML generations and called it the best model she has tested.
Deep code review is the other gap. On Theo's own bench for finding real improvements in a large codebase, Grok 4.7 and GPT-6 Astra are more thorough. Opus 5.5 nearly doubled Opus 5's accuracy there, but OpenAI's frontier still leads — a task he does every couple of weeks, not daily.
3D and game generation, by contrast, impressed him and others: a Dark Souls clone, a from-scratch C++ Tony Hawk-style skater, flight sims, and his own Fish Slop demo running at a smooth 120 FPS with an actual gameplay loop. It seems stronger in three.js in the browser than in native code.
Remaining quirks
The model is smaller and spikier than Fable, and its failure modes show up in long sessions. The most irritating is context-window paranoia: Theo caught an instance reporting that another agent's context was "10% from autocompact" and worrying that a crash would lose work — nonsense concerns the model should not be entertaining, and a behavior OpenAI has trained out of its models. He saw similar-to-lower rates of severe issues in PRs than Fable, with slightly more nitpicks, on only a day of data.
His summary: not a Fable killer on every axis, but the right default for most work, and by far the best value per subscription dollar right now.